On the Generalization Ability of On

Transcription

On the Generalization Ability of On
2050
IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 50, NO. 9, SEPTEMBER 2004
On the Generalization Ability of On-Line Learning
Algorithms
Nicolò Cesa-Bianchi, Alex Conconi, and Claudio Gentile
Abstract—In this paper, it is shown how to extract a hypothesis
with small risk from the ensemble of hypotheses generated by an
arbitrary on-line learning algorithm run on an independent and
identically distributed (i.i.d.) sample of data. Using a simple large
deviation argument, we prove tight data-dependent bounds for the
risk of this hypothesis in terms of an easily computable statistic
associated with the on-line performance of the ensemble. Via
, we then obtain risk tail bounds
sharp pointwise bounds on
for kernel Perceptron algorithms in terms of the spectrum of the
empirical kernel matrix. These bounds reveal that the linear hypotheses found via our approach achieve optimal tradeoffs between
hinge loss and margin size over the class of all linear functions, an
issue that was left open by previous results.
A distinctive feature of our approach is that the key tools for
our analysis come from the model of prediction of individual sequences; i.e., a model making no probabilistic assumptions on the
source generating the data. In fact, these tools turn out to be so
powerful that we only need very elementary statistical facts to obtain our final risk bounds.
Index Terms—Kernel functions, on-line learning, pattern recognition, perceptron algorithm, statistical learning theory.
I. INTRODUCTION
We allow a learning algorithm to output hypotheses of the
, where is a decision space not necessarily
form
equal to . The goodness of hypothesis on example
is
, where
is a
measured by the quantity
nonnegative and bounded loss function. For instance, in binary
,
, and
classification we may take
(the absolute loss function).
We analyze learning algorithms within the framework of
statistical pattern recognition (see, e.g., [1]). In this framework,
are generated by independent draws
all the examples
from a fixed and unknown probability distribution on
.
This assumption allows us to view the training set as a statistical sample, and thus, to investigate the learning problem as a
problem of statistical inference. In what follows, probabilities
and expectations will be understood with respect to the
fixed and unknown underlying distribution according to which
all examples are drawn. Random variables are denoted in upper
case and their realizations in lower case. We use to denote
the pair
of random variables
and
taking values
and , respectively, We also write the training set as a
in
vector-valued random variable
W
E view pattern recognition as a problem of learning from
, where
examples. An example is a pair
(which we call instance) is a data element and
is the label
associated with it. Instances are tuples of numerical and/or
symbolic attributes. Labels belong to a finite set of symbols
(the class elements) or to an interval of the real line, depending
on whether the task is classification or regression.
A learning algorithm for pattern recognition (or learner for
short) takes in input a training set, i.e., a multiset of examples
, and outputs a function
, assigning a label from to each element of . We call hypothesis
any function returned by a learner. Learning algorithms generate hypotheses from specific families of functions, such as
linear-threshold functions, decision trees, or multivariate polynomials. We call hypothesis space (denoted by ) the family
from which a learning algorithm picks its hypotheses.
Manuscript received March 4, 2003; revised February 17, 2004. This work
was supported in part by the European Commission under the KerMIT Project
IST-2001-25341. The material in this paper was presented in part at the Conference on Advances in Neural Information Processing Systems, Vancouver, BC,
Canada, Dec. 2001. An extended abstract of this paper appeared in Advances in
Neural Information Processing Systems 14. Cambridge, MA: MIT Press, 2002,
pp. 359–366.
N. Cesa-Bianchi and A. Conconi are with DSI, the University of Milan, 20135
Milan, Italy.
G. Gentile is with the Dipartimento di Informatica e Comunicazione, University of Insubria, 21100 Varese, Italy.
Communicated by A. E. Nobel, Associate Editor for Nonparametric Estimation, Classification, and Neural Networks.
Digital Object Identifier 10.1109/TIT.2004.833339
In general, we would like a hypothesis to predict well on examples drawn from the same source that generated the training
set for . In statistical pattern recognition, the success of a hypothesis is measured by the risk of , denoted by
.
drawn from
This is the expected loss of on an example
. We identhe underlying distribution,
tify a generic learner with the random hypothesis
returned by when
is fed in input. Our goal is to keep the
risk of small on most sample realizations, that is,
for small enough
and
. Here the probability is
.
taken with respect to the distribution of the training sample
To achieve this goal, we can use the method of uniform convergence, whose study was pioneered by Vapnik and Chervobe the emnenkis (VC) [2] (see also [3], [4]). Let
pirical risk of on a sample
Uniform convergence means that, for all probability distributions, the empirical risk of is, with high probability, close to
. In the case of a class of
its true risk uniformly over all
0018-9448/04$20.00 © 2004 IEEE
CESA-BIANCHI et al.: ON THE GENERALIZATION ABILITY OF ON-LINE LEARNING ALGORITHMS
-valued functions, a sufficient (and also necessary) condition for uniform convergence is that has finite VC dimension
—similar conditions apply to multivalued or real-valued function classes and bounded losses. If this condition holds then, for
and sample size
each
(1)
holds (for a proof of this result see, e.g., [5]).
can be learned by the
Uniform convergence implies that
empirical risk minimizer, i.e., by the algorithm returning the
hypothesis
Once we have a uniform convergence result like (1), the risk
is immediate. Let
.
analysis for
Then, with probability at least
where we applied (1) in the first and the last step.
A different approach to uniform convergence, pioneered in
[6], replaces the square-root term in (1) with the random quantity
where
is a sample statistic depending on
and is a
universal constant. For example,
can be the empirical
VC entropy [7], [8], the Rademacher complexity [9], or the
maximum discrepancy [10] of the class . In general, this
is signifiapproach is advantageous when the mean of
cantly smaller than
and when large deviations of
are unlikely. In these cases, such “data-dependent” uniform
convergence bounds are stronger than those based on the VC
dimension since, with high probability, we have
In some cases, the statistic
directly depends on the emunder consideration,
pirical behavior of the hypothesis
yielding bounds of the form
(2)
Prominent examples of this kind are the bounds for
depends on the margin
linear-threshold classifiers, where
of [11]–[14], and the bounds for Bayesian mixtures, where
depends on the Kullback–Leibler divergence between the
2051
data-dependent mixture coefficients and the a priori coefficients [15]. Note that bounds of the form (2) leave open
the algorithmic problem of finding the hypothesis
optimizing the tradeoff between the terms
and
.
The techniques based on uniform convergence, which we
have seen so far, lead to probabilistic statements which hold
simultaneously for all hypotheses in the class. A different
approach used to derive data-dependent risk bounds yields
statements which only hold for the hypotheses generated by
learning algorithms satisfying certain properties. Examples
along these lines are the notions of self-bounding learners [16],
[17], algorithmic stability [18], and algorithmic luckiness [19].
In this paper, we follow a similar idea and develop a general
framework for analyzing the risk of hypotheses generated
by on-line learners, a specific class of learning algorithms.
Exploiting certain properties of on-line learners, we prove new
data-dependent results via elementary large deviation theory
(Section II), thus avoiding the sophisticated statistical tools
required by risk analyses based on uniform convergence. We
borrow results from the literature on prediction of individual sequences (see, e.g., [20]–[24] for early references on the subject,
and [25]–[31] for specific work on the pattern classification
problem). Based on strong pointwise bounds on the sample
statistic governing the risk for a specific on-line learner, the
kernel Perceptron algorithm, we derive sharp tail risk bounds
for linear hypotheses in terms of the spectrum of the empirical
kernel (Gram) matrix (Section III). Though our bounds are
not easily comparable with previous spectral bounds based on
uniform convergence, there are two reasons for which ours
might be preferable: The empirical kernel matrices occurring
in our bounds are sparse, as they only contain a subset of
“support” examples; moreover, the linear hypotheses found
via our approach achieve optimal tradeoffs between empirical
hinge loss and margin size over the class of all linear functions,
thus solving the algorithmic problem left open by results like
(2).
II. RISK ANALYSIS FOR ON-LINE ALGORITHMS
Unlike standard learning algorithms, on-line learners take in
input a hypothesis
and an example
and return a
(on empty inputs, a default hypothesis
new hypothesis
is returned). We now illustrate how an on-line algorithm is
. First, uses
run on a sample
the first example
and the default hypothesis to gen. Note that
is a random element
erate the first hypothesis
as it depends on the random example
. Next, it
of
and the first hypothesis
uses the second example
to generate the second hypothesis
, and so on. At the end of
of (not necthis process we obtain a sequence
and each
is
essarily distinct) hypotheses, where
and
.
obtained from
of
Our goal is to use the ensemble
hypotheses generated by an on-line learner
to obtain a
,
hypothesis with low risk. An obvious choice is to pick
.
i.e., the hypothesis that depends on the whole training set
However, without making specific assumptions on the way
2052
IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 50, NO. 9, SEPTEMBER 2004
operates, we cannot say about
much more than what could
be said based on standard uniform convergence arguments. In
what follows, we propose a different approach which derives a
(we actually
hypothesis from the ensemble
, even though we do so for purely
discard the last hypothesis
technical reasons).
and an on-line learner , we call
Given a sample
the ensemble of hypotheses generated by . A
central role in our analysis is played by the sample statistic
The statistic
can be easily computed as the on-line algorithm
measures how good on average each
did
is run on .
on the next example
. In agreement with this intuition,
is close to the ensemble’s average risk.
we now prove that
Since we are working with bounded losses, there is no loss of
generality in assuming that the loss function satisfies
The assumption on ensures that
is indeed a map from
to . It is now fairly easy to prove a data-dependent bound on
the risk of .
Corollary 2: Let
be the ensemble of hypotheses generated by an arbitrary on-line algorithm working
with a loss function which is convex in its first argument and
. Then, for any
satisfying
Proof: Using Jensen inequality and the linearity of expectation
Proposition 1: Let
be the ensemble of hypotheses generated by an arbitrary on-line algorithm working
. Then, for any
with a loss satisfying
An application of Proposition 1 concludes the proof.
As a remark, we note that a version of Corollary 2 restricted
to the absolute loss was given in [33].
Proof: For each
set
We have
Furthermore,
. Finally, using
holds since takes values in
to denote the conditional expectation
A direct application of the Hoeffding–Azuma inequality (a generalization of Chernoff–Hoeffding bounds to sums of conditionally zero-mean bounded random variables [32]) to the random
proves the lemma.
variables
B. Risk Analysis for General (Bounded) Losses
Proposition 1 tells us that at least one element in the ensemble
has risk not larger than
with high probability. In this subsection, we show how to identify such a hypothesis with high
probability and with no conditions on the loss function other
than boundedness. Our approach is related to a result of Littlestone [34]. However, unlike Littlestone’s, our technique does
not require a cross-validation set.
The main idea for finding a hypothesis in the ensemble whose
risk is close to the ensemble’s average risk is to compute the emon the sequence of remaining
pirical risk of each hypothesis
examples. Then, to compensate for the fact that the hypotheses
have been tested on portions of the sample of different length,
a different penalization term is added to the empirical risk of
each .
Formally, define the penalized empirical risk of hypothesis
by
where
We will be using this simple concentration result several
times in the rest of this section.
A. Risk Analysis for Convex Losses
Using Proposition 1, it is now possible to derive a data-dependent risk bound of the same form as the one mentioned in
Section I. Suppose that the decision space of is a convex
set and that the loss function is convex in its first argument.
Define the average hypothesis
is the empirical risk of
and
on the remaining sample
is a penalization function. Note that the penalized empirical risk
is a sample statistic. Note also that the penalization function explicitly depends on the confidence parameter . This somewhat
unusual dependence comes from the need of matching the size
CESA-BIANCHI et al.: ON THE GENERALIZATION ABILITY OF ON-LINE LEARNING ALGORITHMS
2053
of the confidence interval, at confidence level
, provided
by the Chernoff–Hoeffding bound.
minimizing the penalOur learner returns the hypothesis
ized risk estimate over all hypotheses in the ensemble, i.e.,
(3)
Our proof technique builds on the concentration result contained
in Proposition 1, and proceeds essentially in two steps. We first
is close to the minimal penalized
prove that the true risk of
risk over the random hypotheses in the ensemble (Lemma 3);
then we exploit the fact that, simultaneously for all these random
hypotheses, the true risk is close to the corresponding empirical
risk (Theorems 4 and 5).
The proof of the next lemma is given in the appendix.
Lemma 3: Let
be the ensemble of hypotheses
generated by an arbitrary on-line algorithm working with a
. Then, for any
, the
loss satisfying
satisfies
hypothesis
where the last inequality follows from
.
Now, it is clear that Proposition 1 can be immediately
generalized to the following set of inequalities, one for each
:
The following theorem is our main result.
Theorem 4: Let
be the ensemble of hypotheses generated by an arbitrary on-line algorithm working
. Then, for any
, the
with a loss satisfying
minimizing the penalized empirical risk based
hypothesis
on
(i.e., the one obtained by replacing with
in (3))
satisfies
(4)
Hence, setting for brevity
we can write
As a matter of fact, we will be proving the tighter statement
contained in the following Theorem 5. The bound in this theorem is formally similar to model selection inequalities, such as
those in [10]. Theorem 4 is obtained by replacing the minimum
in the statement of Theorem 5 with
.
over
Theorem 5: Under the assumptions of Theorem 4, we have
that
is at most , where
Proof: Applying Lemma 3 with
we obtain
where in the last inequality we used (4).
Combining with Lemma 3, we have that with probability at
least
We then observe that
thereby concluding the proof.
It should be clear that both Theorems 4 and 5 could be used to
prove generalization bounds for specific algorithms. In fact, for
2054
IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 50, NO. 9, SEPTEMBER 2004
the sake of comparison to the existing literature, in the rest of
this paper we will be essentially employing Theorem 4. Needless to say, one could recast our results (Theorems 6 and 7 in
Section III) in the style of Theorem 5, instead.
III. APPLICATIONS
As shown in Section II-B, we know how to extract, from the
of hypotheses generated by an arensemble
bitrary on-line learner , a hypothesis whose risk satisfies
with norm defined by
If this space is completed, by adding all limit points of sequences
that are convergent in the norm
, the reis called reproducing kernel Hilbert space.
sulting space
The hypothesis space for the kernel Perceptron algorithm, run
for
.
with kernel , is the space of functions
is the constant
funcThe algorithm’s initial hypothesis
is defined by
tion. The th hypothesis
with high probability, where
and is a universal constant. We now show applications of
this bound to kernel-based linear algorithms Using pointwise
(i.e., bounds on
holding for every posbounds on
of the sample
) we obtain sharp tail risk
sible realization
bounds in terms of the spectral properties of the empirical kernel
Gram matrix generated during the algorithm’s run.
In the sequel we will be focusing on binary classification
tasks with linear-threshold learners. A loss function naturally
associated to such tasks is the – loss, which we will be handling through Theorem 4. As a matter of fact, analogs of our results exist, e.g., for linear regression problems with square loss.
Clearly, for such convex losses we could use Corollary 2, instead.
and
For binary classification tasks we take
. The loss function is the standard – loss:
if
and
otherwise. On-line linear-threshold
learners generate hypotheses of the form1
,
is a so-called weight vector associated with hywhere
pothesis .
We begin by considering the classical Perceptron algorithm
[35]–[37] in its dual kernel form, as investigated in, e.g., [27],
[38]. For an introduction to kernels in learning theory the reader
is referred to [39], or to [40] for a more in-depth monography.
Here, we just recall the following basic definitions.
satisA kernel is a nonnegative function
fying
where
is the set of indices of previously mistaken examples,
and
.
i.e., the set of all such that
We will now prove a specialization of Theorem 4 to the kernel
for the
Perceptron algorithm by using a known bound on
ensemble of hypotheses generated by this algorithm. The bound
uses a loss function different from the – loss. Define
on
at margin
(see, e.g.,
the hinge loss
[27], [28]) by
(5)
Note that when using linear threshold functions the hinge loss is
,
an upper bound on the – loss; i.e.,
,
, and
. For any function
for any
and sample
define
We are now ready to state and prove the risk bound for the kernel
Perceptron algorithm.
Theorem 6: Let
be the ensemble of hypotheses generated by the kernel Perceptron algorithm using
, the hypothesis
minikernel . Then, for any
mizing the penalized empirical risk based on
(i.e., the one
in (3)) satisfies
obtained by replacing with
(6)
with probability at least
for all
,
, and
(such
functions are also called positive definite). Given a kernel, we
can define the linear space
, where2
Proof: For each sample realization , we upper-bound
using known generalizations of the Perceptron convergence theorem (see, e.g., [26], [29]). An application of Theorem 4 concludes the proof.
It is interesting to compare the above bound with the result
obtained in [9, Theorem 21] by Bartlett and Mendelson using
1Here
and in what follows we use (1)
(1) to denote the signum function.
to denote vector transposition and
2From an algorithmic standpoint, it should be observed that the minimum of
(3) can only be achieved on indices 2 M. The hypothesis obtained from the
kernel Perceptron algorithm can then be computed by counting only the number
of mistaken examples of the hypotheses
such that 2 M [ f0g.
t
H
t
H
CESA-BIANCHI et al.: ON THE GENERALIZATION ABILITY OF ON-LINE LEARNING ALGORITHMS
2055
uniform convergence techniques. They prove that, for all fixed
standard (first-order) Perceptron algorithm. Using a pointwise
proven in [26], we can show the following result,
bound on
which uses the same notation as the statement of Theorem 6.
This bound allows us to replace the trace of the kernel Gram
matrix occurring in both (6) and (7) with more detailed spectral
information.
(7)
Theorem 7: Let
be the ensemble of hypotheses generated by the kernel second-order Perceptron
and run with input parameter
.
algorithm using kernel
Then, for any
, the hypothesis
minimizing the
satisfies (8) at the
penalized empirical risk based on
, where
bottom of the page, with probability at least
are the (random) eigenvalues of the kernel Gram
matrix including only those instances
such that
.
holds with probability at least
simultaneously for all functions in the class of all (random) functions of the form
and such that
is similar to (5), but is based on a loss
. The quantity
defined by
if
otherwise.
While comparing (6) with (7), one should observe that bound (7)
holds for all functions of the above form; moreover, the hinge
loss terms obey the inequality
. On
the other hand, the fact that (6) holds for a specific function
bears some crucial advantages over (7). In fact, it solves the
optimizing
algorithmical issue of finding the function
the choice of . As a consequence, our bound does not have
).
a size parameter (which has to be traded off against
Moreover, the main square-root term of our bound only contains
the trace (i.e., the sum of the eigenvalues) of a small submatrix
of the kernel Gram matrix.
The bound of Theorem 6 can be improved by using a recently
proposed variant [26] of the kernel Perceptron algorithm. This
variant, called (kernel) second-order Perceptron algorithm, genof the following form. Fix
erates ensembles of hypotheses
and, as before, let
a sample realization
, the set of indices of
previously mistaken examples. Then
where
•
is the
sparse kernel Gram matrix with
for
;
entries
identity matrix;
• is the
•
is the vector of elements , for
;
is the vector of elements
, for
;
•
• is a nonnegative parameter.
We remark that the inverse matrix
does not have to
be recomputed at each step. Indeed, standard techniques allow
to incrementally update the inverse in time linear in the size of
the matrix.
For a detailed analysis of this algorithm we refer the reader
to [26]. Here we just note that for large values of , the secondorder Perceptron algorithm approximates the behavior of the
A simple comparison between (6) and (8) might go as
follows (the reader is referred to [26] for a more detailed
argument). First, notice that the two bounds only differ in the
kernel/eigenvalue terms under the square root sign. Now, the
in (6) is actually the trace of the kernel
sum
with
. In
Gram matrix including only the instances
.
turn, this trace equals the sum of the eigenvalues
Therefore, (6) has a linear dependence on the (random) eigenvalues , while (8) has only a logarithmic dependence. The
price we pay for this logarithmic dependence is the further
. Now, if a function achieving a
factor
tends to be aligned to an eigenvector of the
small value of
kernel Gram matrix with a small eigenvalue, then this factor
tends to be small, too. In fact, if is perfectly aligned with an
is exactly
eigenvector with eigenvalue , then
equal to . Hence, if parameter is suitably chosen, then the
product
occurring in (8) can be much smaller then the trace
occurring in (6).
Other examples of risk bounds involving the spectrum of the
kernel Gram matrix have been proven in [8 (see Theorem 5.2
therein)] via uniform convergence techniques. These bounds are
not readily comparable to ours. In fact, the results in [8] are
proven through covering numbers arguments which consider the
kernel Gram matrix of the whole sample, whereas our Theorem 7 considers instead a submatrix of including only instances
with
. On the other hand, the results in [8] are
expressed in terms of the “large” eigenvalues of only—where
“large” is defined in terms of the margin of the data, whereas
our bounds are in terms of all the eigenvalues of this submatrix
(note, however, that the eigenvalues of this submatrix cannot
be larger than the corresponding eigenvalues of ; rather, they
are usually quite smaller). Furthermore, unlike our bounds, the
(8)
2056
IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 50, NO. 9, SEPTEMBER 2004
bounds proven in [8] appear to rely on the assumption that there
achieving zero risk.
exists some
Further data-dependent bounds for kernel-based linear algorithms are derived in [11], [18], [19]. These bounds are not expressed in terms of the spectrum of the kernel matrix, and thus
are not easily comparable to ours.
IV. CONCLUSION AND OPEN PROBLEMS
In this paper, we have shown how to obtain sharp data-dependent tail bounds on the risk of hypotheses generated by on-line
learning algorithms. The key analytical tools used in the proofs
are borrowed from the model of prediction of individual sequences, an on-line learning model making no stochastic assumptions on the way the sequence of examples is generated.
Surprisingly, these tools turn out to be so powerful that we only
need very elementary statistical inequalities to complete the argument and obtain our final risk bounds.
It is interesting to note that any learning algorithm can be
turned into an on-line algorithm (possibly at the expense of computational efficiency) by rerunning it on the entire sample after
each arrival of a new example. Hence, our results of Section II
can be used to obtain data-dependent risk bounds, expressed in
, for any learning algorithm. In
terms of the sample statistic
this respect, the advantage of genuinely on-line algorithms is the
, derived by exploiting the
existence of accurate bounds on
structure of the on-line learning process. No such bounds are
known for batch algorithms.
All the results we have presented have convergence rates of
. One might wonder whether it is possible, in our
the form
when
happens to be small.
framework, to achieve rates
Indeed, a more careful use of large-deviation tools (along with
a more involved analysis) allows to sharpen the results we have
given here. As a simple example, the bound of Corollary 2 could
be replaced by
and let
Note that
defined in (3) coincides with
tion, and since
certainty, we can write
. With this notaholds with
(9)
where is a positive-valued random variable whose value will
be specified later. Now, if
holds, then at least one of the following three conditions:
must hold. Hence, for any fixed we can write
(10)
This bound has asymptotical rate
whenever the cumulative
of the underlying on-line
loss
algorithm remains bounded. Though such sharper bounds are
very interesting in their own right, we believe they do not add
much to the overall approach of this paper in terms of simplicity
and conciseness; for this reason, these bounds will be investigated in a future paper.
We close by mentioning an important open question: Is it possible to extend the results of Section II to the case when the examples are not independent; e.g., when they are generated by a
stationary process?
APPENDIX
PROOF OF LEMMA 3
Let
and let
. Set, for brevity
(11)
(12)
Probability (12) is zero if
(10) and (11) into (9) we can write
. Hence, plugging
CESA-BIANCHI et al.: ON THE GENERALIZATION ABILITY OF ON-LINE LEARNING ALGORITHMS
where in the last two inequalities we applied Chernoff–Hoeffding bounds (see, e.g., [1, Ch. 8]) to the random variables
with mean
.
ACKNOWLEDGMENT
The authors wish to thank the anonymous reviewers, and also
the Associate Editor, for their many useful comments.
REFERENCES
[1] L. Devroye, L. Gy˝orfi, and G. Lugosi, A Probabilistic Theory of Pattern
Recognition. Berlin, Germany: Springer-Verlag, 1996.
[2] V. Vapnik and A. Chervonenkis, “On the uniform convergence of relative
frequencies of events to their probabilities,” Theory Probab. Its Applic.,
vol. 16, no. 2, pp. 264–280, 1971.
[3] V. Vapnik, “Inductive principles of the search for empirical dependencies,” in Proc. 2nd Annu. Workshop on Computational Learning Theory,
1989, pp. 3–21.
, The Nature of Statistical Learning Theory, 2nd ed. Berlin, Ger[4]
many: Springer-Verlag, 1999.
[5] P. Long, “The complexity of learning according to two models of a
drifting environment,” Machine Learning, vol. 37, no. 3, pp. 337–354,
1999.
[6] P. Bartlett, “The sample complexity of pattern classification with neural
networks,” IEEE Trans. Inform. Theory, vol. 44, pp. 525–536, Mar.
1998.
[7] S. Boucheron, G. Lugosi, and P. Massart, “A sharp concentration inequality with applications,” Random Structures and Algorithms, vol. 16,
pp. 277–292, 2000.
[8] R. Williamson, J. Shawe-Taylor, B. Schölkopf, and A. Smola, “Sample
based generalization bounds,” NeuroCOLT, Tech. Rep. NC-TR-99-055,
1999.
[9] P. Bartlett and S. Mendelson, “Rademacher and Gaussian complexities:
Risk bounds and structural results,” J. Machine Learning Res., vol. 3,
pp. 463–482, 2002.
[10] P. Bartlett, S. Boucheron, and G. Lugosi, “Model selection and error
estimation,” Machine Learning, vol. 48, pp. 85–113, 2001.
[11] A. Antos, B. Kégl, T. Linder, and G. Lugosi, “Data-dependent marginbased generalization bounds for classification,” J. Machine Learning
Res., vol. 3, pp. 73–98, 2002.
[12] V. Koltchinskii and D. Panchenko, “Empirical margin distributions and
bounding the generalization error of combined classifiers,” Ann. Statist.,
vol. 30, no. 1, pp. 1–50, 2002.
[13] J. Langford, M. Seeger, and N. Megiddo, “An improved predictive accuracy bound for averaging classifiers,” in Proc. 18th Int. Conf. Machine
Learning, 2001, pp. 290–297.
[14] R. Schapire, Y. Freund, P. Bartlett, and W. Lee, “Boosting the margin: A
new explanation for the effectiveness of voting methods,” Ann. Statist.,
vol. 26, no. 5, pp. 1651–1686, 1998.
[15] R. Meir and T. Zhang, “Generalization error bounds for Bayesian mixture algorithms,” J. Machine Learning Res., vol. 4, pp. 839–860, 2003.
[16] A. Blum and J. Langford, “Microchoice bounds and self bounding
learning algorithms,” Machine Learning, vol. 51, no. 2, pp. 165–179,
2003.
2057
[17] Y. Freund, “Self bounding learning algorithms,” in Proc. 11th Annu.
Conf. Computational Learning Theory, 1998, pp. 127–135.
[18] O. Bousquet and A. Elisseff, “Stability and generalization,” J. Machine
Learning Res., vol. 2, pp. 499–526, 2002.
[19] R. Herbrich and R. Williamson, “Algorithmic luckiness,” J. Machine
Learning Res., vol. 3, pp. 175–212, 2002.
[20] N. Cesa-Bianchi, Y. Freund, D. Helmbold, D. Haussler, R. Schapire, and
M. Warmuth, “How to use expert advice,” J. Assoc. Comput. Mach., vol.
44, no. 3, pp. 427–485, 1997.
[21] M. Feder, N. Merhav, and M. Gutman, “Universal prediction of individual sequences,” IEEE Trans. Inform. Theory, vol. 38, pp. 1258–1270,
July 1992.
[22] D. Haussler, J. Kivinen, and M. Warmuth, “Sequential prediction of individual sequences under general loss functions,” IEEE Trans. Inform.
Theory, vol. 44, pp. 1906–1925, Sept. 1998.
[23] N. Littlestone and M. Warmuth, “The weighted majority algorithm,” Inform. Comput., vol. 108, pp. 212–261, 1994.
[24] V. Vovk, “A game of prediction with expert advice,” J. Comp. Syst. Sci.,
vol. 56, no. 2, pp. 153–173, 1998.
[25] P. Auer and M. Warmuth, “Tracking the best disjunction,” Machine
Learning, vol. 32, no. 2, pp. 127–150, 1998.
[26] N. Cesa-Bianchi, A. Conconi, and C. Gentile, “A second-order perceptron algorithm,” in Proc. 5th Annu. Conf. Computational Learning
Theory (Lecture Notes in Artificial Intelligence). Berlin, Germany,
2002, vol. 2375, pp. 121–137.
[27] Y. Freund and R. Schapire, “Large margin classification using the perceptron algorithm,” Machine Learning, pp. 277–296, 1999.
[28] C. Gentile and M. Warmuth, “Linear hinge loss and average margin,” in
Advances in Neural Information Processing Systems 10. Cambridge,
MA: MIT Press, 1999, pp. 225–231.
[29] C. Gentile, “The robustness of the p-norm algorithms,” Machine
Learning, vol. 53, no. 3, pp. 265–299, 2003.
[30] A. Grove, N. Littlestone, and D. Schuurmans, “General convergence results for linear discriminant updates,” Machine Learning, vol. 43, no. 3,
pp. 173–210, 2001.
[31] N. Littlestone, “Redundant noisy attributes, attribute errors, and linear
threshold learning using winnow,” in Proc. 4th Annu. Workshop on Computational Learning Theory, 1991, pp. 147–156.
[32] K. Azuma, “Weighted sums of certain dependent random variables,” Tohoku Math. J., vol. 68, pp. 357–367, 1967.
[33] A. Blum, A. Kalai, and J. Langford, “Beating the hold-out: Bounds
for k -fold and progressive cross-validation,” in Proc. 12th Annu. Conf.
Computational Learning Theory, 1999, pp. 203–208.
[34] N. Littlestone, “From on-line to batch learning,” in Proc. 2nd Annu.
Workshop on Computational Learning Theory, 1989, pp. 269–284.
[35] H. Block, “The perceptron: A model for brain functioning,” Rev. Mod.
Phys., vol. 34, pp. 123–135, 1962.
[36] A. Novikoff, “On convergence proofs of perceptrons,” in Proc. Symp.
Mathematical Theory of Automata, vol. XII, 1962, pp. 615–622.
[37] F. Rosenblatt, “The perceptron: A probabilistic model for information
storage and organization in the brain,” Psychological Rev., vol. 65, pp.
386–408, 1958.
[38] M. Aizerman, E. Braverman, and L. Rozonoer, “Theoretical foundations
of the potential function method in pattern recognition learning, automation and remote control,” Automat. Remote Contr., vol. 25, pp. 821–837,
1964.
[39] N. Cristianini and J. Shawe-Taylor, An Introduction to Support Vector
Machines. Cambridge, U.K.: Cambridge Univ. Press, 2001.
[40] B. Schölkopf and A. Smola, Learning With Kernels. Cambridge, MA:
MIT Press, 2002.